doc2dial dataset is a goal-oriented document-grounded dialogue dataset.
For more details, please refer to our  PAPER ,  README  and  CODE . You can find previous dataset (v0.9)  here .

Data Download

Doc2Dial Data Construction

span labeling
span linking
dialogue flow generating
(1) grounding span
(2) role
(3) da
(4) utterance
dialogue scene
document data
dialogue data
 document processing

Document Data

Each document is annotated with domain, title , document content and span information.

Dialogue Data

Each dialogue is annotationed with document id, domain. Each dialogue turn is annotated with a dialogue scene, which includes role, da and grounding reference. Please note that we also roughly label the span content as either conditional context ("condition") or relevant solution ("solution") for the purpose of designing the dialogue scene. Please see table below for the description of dialogue scene for creating corresponding utterance via crowdsourcing.

Scenes Note: the dialog act annotations are based on Doc2Dial v0.9. We have simplified DA annotations for v1.0.
Dialogue Act Role Description
request/query/open user Ask a question regarding the text (condition) as an issue or request that someone need help with.
respond/yes/agree user The agent just asked a question regarding whether the text (condition) applies to someone, please confirm it.
respond/no/disagree user The agent just asked a question regarding whether the text (condition) applies to someone, please deny it.
assert/provide/precondition_pos user Describe the text (condition) as an issue or request that someone needs help with. Or confirm it when need to answer agent's question.
assert/provide/precondition_neg user Describe the text (condition) as NOT applicable or related to someone. Or deny it when need to agent's question.
assert/provide/solution-neg user Describe the text (solution) as NOT helpful or related to someone.
request/query/open user Ask a question that can be answered with the text (solution).
request/query/open agent Ask a question that aims to verify whether the text (condition) applies to someone.
respond/reply/open agent Response based on the text (solution).
respond/reply/open_pos agent Positively response based on the text (solution), i.e., describe it as applicable to user's query/request.
respond/reply/open_neg agent Negatively response based on the text (solution), i.e., describe it as NOT applicable to user's query/request.
request/query/follow-up user Ask a follow-up question based on the text (solution). The question should be relevant but not the same as previous mentions.
respond/noSolution agent After user providing negative response, respond as there is no solution applicable.
request/query/ood user User provides a query that is irrelevant to the given document, in other words, out of domain (OOD).
respond/reply/ood agent Response with "Irrelevant" to user's OOD query.